Deep features from Pre-trained CNNs and a vision transformer for amazonian timber species recognition using classical machine learning and random image patches
Journal of the Indian Academy of Wood Science · 2026-12 · 23 (2) · 542-562
DOI: 10.1007/s13196-026-00425-5 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of the Indian Academy of Wood Science, online erschienen am 01.09.2026. Verfasst von Deivison Venicio Souza und 10 weiteren Autor:innen. Die Arbeit nennt 79 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Universidade Federal Do Pará.
Abstract (Originalsprache)
Accurate identification of timber species is essential for biodiversity conservation, sustainable forest management, and timber supply chain control. This study evaluated four pre-trained convolutional neural networks—ResNet50, VGG16, InceptionV3, and MobileNetV2 and a Vision Transformer (ViT) as fixed feature extractors for the automated identification of Amazonian timber species. The resulting deep representations were classified using an Artificial Neural Network (ANN), Support Vector Classifier (SVC), Linear Discriminant Analysis (LDA), and Random Forest (RF). The dataset comprised 864 macroscopic images of wood cross-sections representing six species and five genera from the Brazilian Amazon. Two image-representation strategies were compared: original images and 20 random 224 × 224-pixel patches extracted from each image. Patch-level class probabilities were aggregated at the image level through soft voting, and predictive performance was evaluated using tree-level Leave-One-Group-Out Cross-Validation. The patch-based strategy improved classifier performance and produced more consistent recognition across species, particularly by increasing the recall of classes with the highest error rates when original images were used. Among the evaluated feature extractors, ViT representations yielded the best results when combined with ANN, SVC, and LDA. The ViT + SVC model achieved the highest predictive performance, with 98.29% accuracy and an MCC of 97.96%. ViT + LDA achieved comparable performance, with 97.70% accuracy and an MCC of 97.26%, while requiring substantially less time for hyperparameter optimization. These findings indicate that combining pre-trained deep feature extractors, conventional machine learning algorithms, random patch sampling, and soft-voting aggregation is a promising and computationally efficient strategy for the automated identification of Amazonian timber species.
Förderung
- Universidade Federal Do Pará
Zitieren
APA
Souza, D. V., Santos, J. X. D., Gama, N. C., Silva, W. D. S. D., Sousa, G. S., Vieira, H. C., Acosta, T. L. N., Cirilo, A. J. C., Silva, G. C. C., Muñiz, G. I. B., & Nisgoski, S. (2026). Deep features from Pre-trained CNNs and a vision transformer for amazonian timber species recognition using classical machine learning and random image patches. Journal of the Indian Academy of Wood Science, 23(2), 542-562. https://doi.org/10.1007/s13196-026-00425-5
Vancouver
Souza DV, Santos JXD, Gama NC, Silva WDSD, Sousa GS, Vieira HC, et al. Deep features from Pre-trained CNNs and a vision transformer for amazonian timber species recognition using classical machine learning and random image patches. Journal of the Indian Academy of Wood Science. 2026;23(2):542-562. doi:10.1007/s13196-026-00425-5
BibTeX
@article{Souza2026,
author = {Souza, Deivison Venicio and Santos, Joielan Xipaia dos and Gama, Natally Celestino and Silva, Washington Duarte Silva da and Sousa, Genilson Santana and Vieira, Helena Cristina and Acosta, Tawani Lorena Naide and Cirilo, Alexia Juliana Camodeco and Silva, Gabriela Cristina Costa and Muñiz, Graciela Inés Bolzon and Nisgoski, Silvana},
title = {Deep features from Pre-trained CNNs and a vision transformer for amazonian timber species recognition using classical machine learning and random image patches},
journal = {Journal of the Indian Academy of Wood Science},
year = {2026},
volume = {23},
number = {2},
pages = {542--562},
doi = {10.1007/s13196-026-00425-5}
}
EUVIMED